train.py

# %%
import random
import numpy as np
import torch
import torch.nn as nn
import torchvision.transforms as transforms
from torch.optim.lr_scheduler import ReduceLROnPlateau
from PIL import Image
from torch.utils.data import ConcatDataset, DataLoader, Dataset
from torchvision.datasets import DatasetFolder
from tqdm import tqdm
from PIL import ImageFile

ImageFile.LOAD_TRUNCATED_IMAGES = True
random.seed(1234)
np.random.seed(1234)
torch.manual_seed(1234)

folder = "./datasets"
NUM_CLASSES = 14


train_tfm = transforms.Compose(
    [
        # Resize the image to a fixed shape (height = width = 256)
        transforms.Resize((224, 224)),
        transforms.Lambda(lambda x: x.convert("RGB")),
        # Random horizontal flip to increase robustness to object orientation
        transforms.RandomHorizontalFlip(),
        # Random rotation, a common transformation to handle rotated images
        transforms.RandomRotation(
            20
        ),  # Rotate image by a random angle between -20 and 20 degrees
        # Random cropping to simulate random scene zoom
        transforms.RandomResizedCrop(
            224, scale=(0.8, 1.0)
        ),  # Crop and resize to 224x224
        # Random color jitter to make the model robust to lighting changes
        # transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.2),
        # Random affine transformation (translation, scaling, rotation)
        # transforms.RandomAffine(degrees=10, translate=(0.1, 0.1), scale=(0.8, 1.2)),
        # Convert to tensor
        transforms.ToTensor(),
        # Normalize the image with mean and standard deviation for better convergence
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]
)

test_tfm = transforms.Compose(
    [
        # Resize to the fixed size
        transforms.Resize((224, 224)),
        transforms.Lambda(lambda x: x.convert("RGB")),
        # Convert to tensor
        transforms.ToTensor(),
        # Normalize the image with mean and standard deviation (same as in training)
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]
)


def get_dataset():
    train_set = DatasetFolder(
        folder + "/train/labeled",
        loader=lambda x: Image.open(x),
        extensions="jpg",
        # transform=train_tfm,
    )
    valid_set = DatasetFolder(
        folder + "/val",
        loader=lambda x: Image.open(x),
        extensions="jpg",
        transform=test_tfm,
    )
    unlabeled_set = DatasetFolder(
        folder + "/train/unlabeled",
        loader=lambda x: Image.open(x),
        extensions="jpg",
        # transform=train_tfm,
    )
    test_set = DatasetFolder(
        folder + "/test",
        loader=lambda x: Image.open(x),
        extensions="jpg",
        transform=test_tfm,
    )
    return train_set, valid_set, unlabeled_set, test_set


def train_collate_fn(batch):
    data, labels = zip(*batch)
    # data = torch.stack(data)
    labels = torch.tensor(labels)
    return data, labels


def test_collate_fn(batch):
    data, labels = zip(*batch)
    data = torch.stack(data)
    labels = torch.tensor(labels)
    return data, labels


from utils import CustomDataset


def update_dataset(
    train_set, unlabeled_set, model, threshold, batch_size=128, num_workers=8
) -> Dataset:
    """
    This is the core function to generate pseudo-labels dataets using the given model.
    inputs:
        - train_set: The labeled training set
        - unlabeled_set: The unlabeled dataset to be pseudo-labeled
        - model: The trained model to generate pseudo-labels
        - threshold: Confidence threshold for pseudo-labeling
        - batch_size: Batch size for DataLoader
        - num_workers: Number of workers for DataLoader
    outputs:
        - new_set: The updated dataset with pseudo-labels
    """
    device = "cuda" if torch.cuda.is_available() else "cpu"

    # Make sure the model is in eval mode.
    model.eval()
    # Define softmax function.
    softmax = nn.Softmax(dim=-1)

    # Create a dataloader for the unlabeled data
    unlabeled_loader = DataLoader(
        unlabeled_set,
        batch_size=batch_size,
        shuffle=False,
        num_workers=num_workers,
        pin_memory=True,
        collate_fn=train_collate_fn,
    )

    # List to store the most confident predictions
    confident_samples = []
    confident_labels = []

    with torch.no_grad():
        for batch_idx, (images, _) in enumerate(
            tqdm(unlabeled_loader, desc="Generating pseudo-labels")
        ):
            # Apply test transform to each image in the batch
            new_images = torch.stack([test_tfm(img) for img in images])

            # Forward pass through the model
            outputs = model(new_images.to(device))

            # Apply softmax to get probabilities
            probabilities = softmax(outputs)

            # Get the maximum probability and corresponding class for each sample
            max_probs, pseudo_labels = torch.max(probabilities, dim=1)

            # For each sample in the batch, check confidence threshold
            for i, (prob, label) in enumerate(zip(max_probs, pseudo_labels)):
                # If the prediction is confident enough, add to confident set
                if prob.item() > threshold:
                    # Get the actual index in the unlabeled_set
                    idx = batch_idx * batch_size + i
                    if idx < len(unlabeled_set):
                        img, _ = unlabeled_set[idx]
                        confident_samples.append(img)
                        confident_labels.append(int(label.cpu()))

    # Create new dataset from the confident predictions
    if confident_samples:
        pseudo_set = CustomDataset(images=confident_samples, labels=confident_labels)

        # Combine with existing labeled data
        new_set = ConcatDataset([train_set, pseudo_set])

        print(f"Added {len(confident_samples)} pseudo-labeled samples to training set")
    else:
        print("No confident pseudo-labels found.")
        new_set = train_set
    return new_set

# %%
from models import *  # noqa: F403
from torch import optim

# "cuda" only when GPUs are available.
device = "cuda:0" if torch.cuda.is_available() else "cpu"

# Initialize a model, and put it on the device specified.
# model = Classifier().to(device)
# model = VGG16Classifier(num_classes=NUM_CLASSES).to(device)
# model = ResNet50Classifier(num_classes=NUM_CLASSES).to(device)
model = ResNet101Classifier(num_classes=NUM_CLASSES).to(device)
# model = InceptionV3Classifier(num_classes=NUM_CLASSES).to(device)
# model = ViTLargeClassifier(num_classes=NUM_CLASSES).to(device)

# !!! ONLY USE THIS FOR RESUME TRAINING !!!
model.load_state_dict(torch.load("best_model.pth"))

criterion = nn.CrossEntropyLoss()

optimizer = optim.Adam(model.parameters(), lr=1e-5, weight_decay=1e-5)

scheduler = ReduceLROnPlateau(
    optimizer,
    mode="min",
    factor=0.1,
    patience=2,
    threshold=0.001,
    threshold_mode="rel",
    cooldown=0,
    min_lr=0,
    eps=1e-08,
)

# %%
# TRAINING
do_semi = True
batch_size = 128
num_workers = 8
train_set, valid_set, unlabeled_set, test_set = get_dataset()

train_loader = DataLoader(
    dataset=train_set,
    batch_size=batch_size,
    shuffle=True,
    num_workers=num_workers,
    collate_fn=train_collate_fn,
)
valid_loader = DataLoader(
    dataset=valid_set,
    batch_size=batch_size,
    shuffle=True,
    num_workers=num_workers,
    collate_fn=test_collate_fn,
)
test_loader = DataLoader(
    dataset=test_set,
    batch_size=batch_size,
    shuffle=False,
    num_workers=num_workers,
    collate_fn=test_collate_fn,
)
best_valid_loss = float("inf")
start_epoch = 100
epochs = 100
threshold = 0.8
early_stop = False
for epoch in range(start_epoch, epochs):
    if do_semi and epoch > epochs // 4 and epoch % 2 == 0:
        new_set = update_dataset(
            train_set=train_set,
            unlabeled_set=unlabeled_set,
            model=model,
            threshold=threshold,
            batch_size=batch_size,
            num_workers=num_workers,
        )
        train_loader = DataLoader(
            dataset=new_set,
            batch_size=batch_size,
            shuffle=True,
            num_workers=num_workers,
            pin_memory=True,
            collate_fn=train_collate_fn,
        )

    # ---------- Training ----------
    model.train()

    train_loss = []
    train_accs = []

    for batch in tqdm(train_loader):
        imgs, labels = batch

        new_images = torch.stack([train_tfm(img) for img in imgs])

        logits = model(new_images.to(device))

        loss = criterion(logits, labels.to(device))

        optimizer.zero_grad()

        loss.backward()

        optimizer.step()

        acc = (logits.argmax(dim=-1) == labels.to(device)).float().mean()

        train_loss.append(loss.item())
        train_accs.append(acc)

    train_loss = sum(train_loss) / len(train_loss)
    train_acc = sum(train_accs) / len(train_accs)

    print(
        f"[ Train | {epoch + 1:03d}/{epochs:03d} ] loss = {train_loss:.5f}, acc = {train_acc:.5f}"
    )
    # ---------- Validation ----------
    model.eval()

    valid_loss = []
    valid_accs = []

    for batch in tqdm(valid_loader):
        imgs, labels = batch

        with torch.no_grad():
            logits = model(imgs.to(device))

        loss = criterion(logits, labels.to(device))

        acc = (logits.argmax(dim=-1) == labels.to(device)).float().mean()

        valid_loss.append(loss.item())
        valid_accs.append(acc)

    valid_loss = sum(valid_loss) / len(valid_loss)
    valid_acc = sum(valid_accs) / len(valid_accs)

    print(
        f"[ Valid | {epoch + 1:03d}/{epochs:03d} ] loss = {valid_loss:.5f}, acc = {valid_acc:.5f}"
    )

    scheduler.step(metrics=valid_loss)

    if valid_loss < best_valid_loss:
        best_valid_loss = valid_loss
        torch.save(model.state_dict(), "best_model.pth")
        print(f"Model saved with loss: {valid_loss:.5f}, acc: {valid_acc:.5f}")
    elif early_stop:
        if epoch > epochs // 2 and valid_loss > best_valid_loss * 1.2:
            print("Early stopping")
            break


# %%
# torch.save(model.state_dict(), "best_model.pth")


# %%
# LOAD BEST MODEL

model.load_state_dict(torch.load("best_model.pth"))
model.eval()


# %%
# TEST

model.eval()

# Initialize a list to store the predictions.
predictions = []


# Iterate the testing set by batches.
for batch in tqdm(test_loader):
    imgs, labels = batch

    with torch.no_grad():
        logits = model(imgs.to(device))

    predictions.extend(logits.argmax(dim=-1).cpu().numpy().tolist())


# %%
# Save predictions into the file.
with open("predict.csv", "w") as f:
    # The first row must be "Id, Category"
    f.write("Id,Category\n")

    # For the rest of the rows, each image id corresponds to a predicted class.
    for i, pred in enumerate(predictions):
        f.write(f"{i},{pred}\n")
print("Predictions saved to predict.csv")